KakaoCloud review, pricing and verdict

KakaoCloud fits teams that need compute and GPUs physically in Korea, from startups to regulated finance. It runs three live Seoul zones in its main region and no region outside Korea. Pick it for local residency and certifications, and look elsewhere if you need GPUs in other countries.

KakaoCloud is the public cloud of Kakao Enterprise, hosted in Seoul. Its pricing overview lists 41 services, including virtual machines, bare metal, managed Kubernetes and GPU instances built on NVIDIA T4 and A100 cards. Billing is pay-as-you-go by the minute. Public institutions have their own regions, and the vendor lists ISMS-P and CSAP certification.

About KakaoCloud

KakaoCloud is the public cloud run by Kakao Enterprise, the enterprise arm of the company behind KakaoTalk. Its pricing overview lists 41 services across compute, networking, storage, containers, databases, analytics, security and AI tooling, all hosted in Seoul. It is built for teams that want Korean data residency and a local support desk rather than a global hyperscaler footprint.

The pricing page frames low cost as the main differentiator. Kakao Enterprise says its rates run up to 15% below a global provider it does not name, based on its own September 2023 comparison, so treat that as a vendor claim. Standard networking pieces are free: VPC, IAM, Cloud Trail, OSS Library and AI Insight (a GPU monitoring service) carry no charge, and inbound traffic is free everywhere.

Compute comes in general-purpose, compute-optimised, memory-optimised and accelerated families. The accelerated line spans three families: NVIDIA T4 cards, FuriosaAI WARBOY NPUs and a higher-end NVIDIA tier sold in one, two or four-card sizes. Kakao Enterprise publishes customer stories on GPU cost, including GS Retail reporting 35% lower GPU spend, which are vendor-supplied figures. Teams comparing GPU-first clouds can read the CoreWeave and Crusoe pages, and teams wanting a plain EU or Asian VPS can compare Hetzner or IDCloudHost.

Billing is pay-as-you-go, charged by the minute from instance creation to deletion, with a monthly figure calculated on 30 days and all published rates excluding VAT. A general-purpose m2a.large with 2 vCPU and 8 GiB costs 102 won per hour, or 73,440 won per month. New accounts receive a 300,000 won credit, and there is no up-front commitment. Outbound traffic is free up to 30 GiB a month and then billed per GiB on a tier that gets cheaper with volume.

The parent company is covered on the Kakao page, which also links Kakao's open-weight Kanana-2 30B-A3B Instruct model. KakaoCloud's pricing pages list infrastructure, not a hosted model API, so that model is a separate record rather than a service you buy here. The older kr-central-1 region was discontinued on 31 March 2025, leaving kr-central-2 as the general region.

Other cloud and GPU hosting pages on Hokai give a wider view of the market: Baseten and Google Vertex AI cover managed model serving, Lium, AtlasCloud and AITECH Cloud Network sell rented GPU capacity, Orion AI Factory targets AI-factory deployments, Hostinger covers budget VPS hosting, and Daytona offers sandboxed dev environments. The directory groups them under GPU compute and cloud.

Kakao has a domestic rival in Naver, whose HyperCLOVA X model and CLOVA X assistant are covered on their own pages.

Pricing

Rates are in Korean won, excluding VAT, billed by the minute, with monthly figures on a 30-day basis. Cheapest paid instance t1i.nano (2 vCPU, 0.5 GiB) is 5.5 won/hour or 3,960 won/month. m2a.large (2 vCPU, 8 GiB) is 102 won/hour or 73,440 won/month; m3az.large is 325 won/hour.

GPU: gn1i.xlarge with one T4 is 648 won/hour; p2i.6xlarge with one A100 is 5,334 won/hour or 3,840,480 won/month; p2i.24xlarge with four A100 is 21,590 won/hour. Kubernetes Engine control plane is 90 won/hour per cluster, with worker nodes billed as VM instances. Object Storage STANDARD is 0.0375 won per GiB-hour up to 1 PiB.

Outbound traffic is free to 30 GiB, then 90 won/GiB to 10 TiB, 80 won/GiB to 20 TiB and 70 won/GiB beyond. Public IP is 5.5 won/hour. Windows Server images add 33 won/hour.

A 300,000 won new-customer credit is advertised.

Key Features

• A100, T4 and NPU instances: The accelerated line offers NVIDIA T4, NVIDIA A100 (1, 2 or 4 cards) and FuriosaAI WARBOY NPU instances, from 648 won per hour for a single T4.
• Per-minute billing: Instance charges run from creation to deletion in minutes, with no commitment and a free credit advertised for new customers.
• Managed Kubernetes at 90 won per hour: Kubernetes Engine charges 90 won per hour for each cluster control plane and bills worker nodes as ordinary VM instances.
• Free core networking and identity: The private network, access control, audit log and GPU monitoring services have no charge, and inbound traffic is free.
• Public-institution regions: Separate regions kr-gov-central-1 and kr-gov-central-2 serve public institution accounts, alongside the general kr-central-2 region with three live zones.
• Korean compliance stack: The vendor lists ISMS-P, CSAP for IaaS, CSA STAR Gold and ISO 27001, 27701, 27017, 27018 and 27799 on its solution pages.

Strengths

• Tiered outbound pricing falls as volume grows and a monthly allowance is free, which suits Korean-facing apps with heavy egress.
• Published per-hour prices for one-, two- and four-card GPU sizes make budgeting possible without a sales call.
• Certification coverage including ISMS-P and CSAP for IaaS matters to Korean finance and public-sector buyers.

Weaknesses

• Everything runs in Seoul: the general region is kr-central-2 only, so teams needing GPUs outside Korea must look elsewhere.
• No service-level percentage was readable on the pages checked, and the SLA terms sit in per-service documents.
• The pages are in Korean, rates are in won and exclude VAT, so overseas teams must convert and add tax themselves.